About min-cut about graphless-neural-networks

[ShichangZh]

Thank you for reaching out. Just to confirm the question here. Are you saying that the first five numbers are the results you get when you apply the min-cut loss to a GNN model? Did you use GraphSAGE, or did you use other architectures?

[ShichangZh]

I see. It is been a while, so I will have to go back and check our code again, which may take some time. Meanwhile, may I ask did you only observe this for Citeseer or other datasets as well? Could you please also provide results on a few more datasets if you have them? Say Cora.

By the way, your previous comment is messed up with some extra email information. You may want to edit it.

[wutaiqiang]

After my doubt check, I find I get such results due to:

I forget to softmax the results for teacher, but I do that for student.

Hence, results from teacher are unstable but from student are not so .

After fix the issue, I get:

teacher:
0.9904
0.9978
0.9898
0.9957
0.9925

student:
0.9457
0.9499
0.9519
0.9670
0.9278

setting: SAGE, Citeseer, and transductive.